Research

Notes from the
workbench.

Working notes, technical write-ups, and longer-form essays from the practice. Published when we have something to say — not on a schedule.

On the discipline of dataset construction: a case for treating data work as design work. Auditability is the feature: a framework for ML systems in regulated industries. Fundus marker detection at clinical resolutions: what we learned building HorusVue. Latency budgets for safety-critical translation: design notes from TIC's first generation. Visual programming for ML pipelines: why the canvas is the right surface for architecture. Distribution drift in longitudinal medical imaging studies: detection and mitigation. What we mean when we say a model has earned trust. A note on ICAO phraseology, NLP, and the limits of generality. The first cut is the dataset: a methodology for early-phase ML engagements. On the careful naming of things: why we call the dataset workbench ALEX.
Stay in touch

An occasional letter from the practice.

Long-form research notes, ~6 to 8 times per year. No marketing, no product announcements unless we have actually built something.